Converting sign language gestures from digital images to text ASL2TXT Converting sign language gestures from digital images to text George Corser
Presentation Overview Concept Foundation: Barkoky & Charkari (2011) Segmentation Thinning My Contribution: Corser (2012) Segmentation (similar to Barkoky) CED: Canny Edge Dilation (Minus Errors) Assumption: User trains his own phone
Concept Deaf and hearing people talking on the phone, each using their natural language Sign-activated commands like voice-activated
Situation: Drive Thru Window Think: Stephen Hawking Deaf person signs order Phone speaks order Confirmation on screen
Process Flow Requires several conversion processes Many have been accomplished Remaining: ASL2TXT
Goal: Find an Algorithm Find an image processing algorithm that recognizes ASL alphabet = A Web site
Barkoky: Segmentation & Thinning Barkoky counts endpoints to determine sign (doesn’t work for ASL)
Barkoky Process Segmentation Thinning Capture RGB image Rescale Extract using colors Reduce noise Crop at wrist Result: hand segment Input: hand segment Apply thinning Find endpoints, joints Calculate lengths Clean short lengths Identify gesture by counting endpoints
1. Capture RGB Image 2. Rescale % ---------- 1. Capture RGB image a = imread('DSC04926.JPG'); figure('Name','RGB image'),imshow(a); % ---------- 2. Rescale image to 205x154 a10 = imresize(a, 0.1); figure('Name','Rescaled image'),imshow(a10);
3. Extract Hand Using Colors % ---------- 3. Extract hand using color abw10 = zeros(205,154,1); for i=1:205, for j=1:154, if a10(i,j,2)<140 && a10(i,j,3)<100, abw10(i,j,1)=255; end; figure('Name','Extracted'),imshow(abw10); Note: Color threshold code differs from Barkoky
Colors: Training Set Histograms
Colors: Training Set (2) Red Green Blue Excel
Colors: Test Set Histograms
4. Reduce Noise % ---------- 4. Reduce noise for i=2:204, for j=1:154, if abw10(i-1,j,1)==0 if abw10(i+1,j,1)==0, abw10(i,j,1)=0; end; end; if abw10(i-1,j,1)==255 if abw10(i+1,j,1)==255, end; end; abw10 = imfill(abw10,'holes');
5. Identify Wrist Position % ---------- 5. Identify wrist position for i=204:-1:1, for j=1:154, if abw10(i,j,1)==255, break; end; end; if j ~= 154 && abw10(i+1,j,1)~=255, wristi=i+1; wristj=j+1; break;
Wrist Detection Algorithm searches bottom-to-top of image Finds a leftmost white pixel above black pixel Sets wrist position SE of found white pixel
Corser: Segmentation & CED Segmentation (similar to Barkoky) Color threshold technique slightly different American Sign Language (ASL) alphabet, not Persian Sign Language (PSL) numbers Image Comparison: Tried Several Methods Full Threshold (Minus Errors) Diced Segments (Minus Errors) Endpoint Count Difference CED: Canny Edge Dilation
ASL Training Set Hit-or-miss: 23% Barkoky: 8%
ASL Test Set MATLAB
A
A
B
B
C
C
D
D
E
E
F
F
G
G
H
H
I
I
J
J
K
K
L
L
M
M
N
N
O
O
P
P
Q
Q
R
R
S
S
T
T
U
U
V
V
W
W
X
X
Y
Y
Z
Z
Z
Hybrid Algorithm Example % ---------- MATLAB Code ------------------- matchtotal = 0; if abs(x10range - x20range) < 20, matchtotal = matchtotal + 10; end; if abs(y10range - y20range) < 20, matchtotal = matchtotal + 11; matchtotal = matchtotal - abs(h10 - h20); % ----- h10, h20 are vector magnitudes -----
Erosion Subtraction
Canny Edge
Canny Edge Dilation Code % ---------- MATLAB Code ------------------- se = strel('disk',5); a10 = edge(a10,'canny'); a20 = edge(a20,'canny'); a10 = imdilate(a10,se); a20 = imdilate(a20,se); % ----- Then calculate matches minus errors
Experimental Results Technique Correct Full Threshold (Minus Errors) 19% (27%) Diced Segments (Minus Errors) 23% (27%) Barkoky Endpoint Count Diff. 8% Hybrid - Height/Width/Endpoints 19% Erosion Subtraction 15% Canny Edge Dilation (Minus Errors) 12% (35%)
Disadvantages Dependent on lighting conditions Fails with flesh-tone backgrounds Requires calibration to a specific user Limited applications: text messaging, activation (“sign” similar to voice activation) ASL numbers (A=10, D=1, O=0, V=2, W=6) Alphabet is tiny portion of full translation: complete translation maybe many years away
Future Work Barkoky claims flesh tones can be detected, but I have yet to replicate (even Barkoky changed his color detection scheme) Could write letter-by-letter algorithm Could use range camera to compute distance of finger instead of shape of hand Motion analysis or edge count Many possibilities… we’ve only just begun! Cue: music http://www.youtube.com/watch?v=__VQX2Xn7tI
The End